Semi-automatic Ground Truth Generation for License Plate Recognition System

摘要

License plate recognition (LPR) system is to help alert relevant personnel of any passing vehicle in the surveillance area. In order to test algorithms for license plate recognition, it is necessary to have input frames in which the ground truth is determined. The purpose of ground truth data is here to provide an absolute reference for performance evaluation or training purposes. However, annotating ground truth data for real-life inputs is very disturbing task because of time-consuming manual. In this paper, we proposed a method of semi-automatic ground truth generation for license plate recognition in video sequences. The method started from region of interesting detection to rapidly extract characters lines followed by a license plate recognition system to verify the license plate regions and recognized the numbers. On the top of the LPR system, we incorporated a tracking-validation mechanism to detect the time interval of passing vehicles in input sequences. The tracking mechanism is initialized by a single license plate region in one frame. Moreover, in order to tolerate the variation of the license plate appearances in the input sequences, the validator would be updated by capturing positive and negatives samples during tracking. Experimental results show that the proposed method can achieve promising results.